Presentation Information
[1Yin-A-25]Overall Quality and Cost Estimation in Generative AI-Driven Development Processes Using Adaptive Subgraph Selection
〇Masatoshi Sekine1, Daisuke Shimbara1 (1. Hitachi, Ltd.)
Keywords:
Software Development,Graph,Reinforcement Learning,Partial Observation,Graph Attention Network
In recent years, while the application of generative AI in software development has been advancing, large-scale development projects face the challenge of enormous token consumption associated with the generation of artifacts such as requirements, specifications, designs, code, and test specifications. This study proposes a method to estimate overall quality and cost with high accuracy by adaptively selecting and generating only a subset of these artifacts. The proposed approach models the entire set of artifacts as a "Generative AI-driven development process graph" with mutual dependencies. It consists of two components: subgraph selection, which sequentially identifies highly significant nodes, and overall evaluation value estimation, which complements unobserved parts based on observed information. Performance evaluations demonstrated that, compared to random selection, the proposed method can estimate overall values with higher accuracy while maintaining the same cost.